ES2608613A1 - Methodology for the automated recognition of reptiles by transforming the markov model of the parametric fusion characteristics of their sound production. (Machine-translation by Google Translate, not legally binding) - Google Patents

Abstract

The present invention relates to a method for the recognition and automated census of reptiles through the transformation using hidden markov models of the fusion of the characteristics in different domains of their acoustic signal emissions allowing the identification of the species and the monitoring specific to individuals within the same species. This methodology uses a concatenation of temporal and spectral characteristics to be automatically recognized by means of an intelligent pattern recognition system. (Machine-translation by Google Translate, not legally binding)

Description

DESCRIPTION

Methodology for the automated recognition of reptiles by transforming the Markov model of the parametric fusion of characteristics of its sound production.

 5

Object of the invention

The present invention relates to a procedure for the recognition and automated census of reptiles through the hyperdimensionality of the transformation of their acoustic signal emissions allowing the identification of the species and the specific monitoring of individuals within the same species. The bio-acoustic signals produced by reptiles are generated in various ways: by excitation of the larynx, expelling air through its nose or mouth, and stirring or scratching body parts among other mechanisms.

Background of the invention

Currently, the use of bio-acoustic techniques for the study and monitoring of animal species within their habitat is one of the most important tools for biologists and conservationists. The technological advance experienced in acoustic sensors and digital recording media allows the census and identification of species remotely avoiding invasive techniques that alter ecosystems or assume the physical presence of the biologist in the study area. The data collected allow the monitoring of animals avoiding their physical marking and provide researchers with information on the biological indicators of the area. The presence or absence of certain species and their number can be used to determine the health of an ecosystem, 25 detecting the presence of pollution, the state of water quality, climatic changes or even alterations in ultraviolet radiation.

There are numerous studies of the spectrum-temporal characteristics of species, in which attempts are made to analyze the parameters in frequency and time of the acoustic signals or vocalizations produced by the animals in order to identify patterns in their communications and their social iterations. In them in general, the procedure consists of collecting hours of sound recordings by means of sensors or microphones located in the study habitat, which are listened to and analyzed spectrum-temporarily by a biologist to determine the presence of a certain species in the area That is being investigated. However, this procedure is slow due to the large number of recording hours that may have been collected and the need to have an expert bio-acoustics biologist familiar with the animal species to which the monitoring is desired. In recent years an effort has been made with the intention of automating this procedure by means of intelligent systems using automatic recognition techniques. Studies have focused on species with extensive sound production such as birds, frogs and whales, where there are several promising investigations that try to solve this problem. They apply techniques used in the recognition of human speech through expert systems that recognize more or less successfully the species under study. On the contrary, reptiles when considered silent or with little sound production have never been the objects of this type of research. However, reptiles including crocodiles, geckos, snakes and turtles are capable of producing bio-acoustic sounds that are specific to the species. The main studies in acoustic recognition have focused on the sounds generated by birds, an example of this can be found in the following articles: 50

i) Harmä, Automatic identification of bird species based on sinusoidal modeling of syllables, in: Acoustics, Speech, and Signal Processing 2003. Proceedings (ICASSP'03). 2003 IEEE International Conference on, Vol. 5, IEEE, 2003, pp. V-545.

ii) S. Fagerlund, Bird species recognition using support vector machines, EURASIP 5 journal on Applied Signal Processing 2007 (1) (2007) 64-64.

iii) Lee, Chang-Hsing, Chin-Chuan Han, and Ching-Chien Chuang. "Automatic classification of bird species from their sounds using two-dimensional cepstral coefficients." Audio, Speech, and Language Processing, IEEE Transactions on 16.8 10 (2008): 1541-1550.

iv) Jančovič, Peter, and Münevver Köküer. "Automatic detection and recognition of tonal bird sounds in noisy environments." EURASIP Journal on Advances in Signal Processing 2011.1 (2011): 982936. 15

v) Graciarena, Martín, et al. "Acoustic front-end optimization for bird species recognition." Acoustics Speech and Signal Processing (ICASSP), 2010 IEEE International Conference on. IEEE, 2010.

 twenty

vi) Graciarena, Martín, et al. "Bird species recognition combining acoustic and sequence modeling." Acoustics, Speech and Signal Processing (ICASSP), 2011 IEEE International Conference on. IEEE, 2011.

vii) Lopes, Marcelo T., et al. "Automatic bird species identification for large number of 25 species." Multimedia (ISM), 2011 IEEE International Symposium on. IEEE, 2011.

viii) Mporas, Iosif, et al. "Automated Acoustic Classification of Bird Species from Real-Field Recordings." Tools with Artificial Intelligence (ICT A 1), 2012 IEEE 24th International Conference on. Vol. L. IEEE, 2012. 30

ix) Juang, Chia-Feng, and Tai-Mou Chen. "Birdsong recognition using prediction-based recurrent neural fuzzy networks." Neurocomputing 71.1 (2007): 121-130.

Classic acoustic automatic recognition techniques have been used for acoustic recognition of patterns, of people and animals, as in:

x) R. Bardelim, Algorithmic analysis of Complex Audio Scenes. Bonn University. PhD Thesis, 2008

 40

xi) H. Xing, P.C. Loizou, Frequency Shift Detection of Speech with GMMs and SVMs, IEEE workshop on Signal Processing Systems, (2002) 215-219

In addition, classic techniques of automatic acoustic recognition of insects, bats and frogs have been tried in the same way, examples of this can be found in the following articles:

xii) K. Riede, Acoustic monitoring of orthoptera and its potential for conservation, Journal of Insect Conservation 2 (3-4) (1998) 217-223.

 fifty

xiii) T. Ganchev, l. Potamitis, N. Fakotakis, Acoustic monitoring of singing insects, in: Acoustics, Speech and Signal Processing, 2007. ICASSP 2007. IEEE International Conference on, Vol. 4, IEEE, 2007, pp. IV-721.

xiv) Z. Leqing, Z. Zhen, lnsect sound recognition based on sbc and hmm, in: Intelligent 5 Computation Technology and Automation (ICICTA), 2010 International Conference on, Vol. 2, IEEE, 2010, pp. 544-548.

xv) D. Chesmore, Automated bioacoustic identification of species, Anais da Brazilian Academy of Sciences 76 (2) (2004) 436-440. 10

xvi) J. Pinhas, V. Soroker, A. Hetzroni, A. Mizrach, M. Teicher, J. Goldberger, Automatic acoustic detection of the red palm weevil, computers and electronics in agriculture 63 (2) (2008) 131-139 .

 fifteen

xvii) A. E. Chaves, C. M. Travieso, A. Camacho, J. B. Alonso, Katydids acoustic classification on verification approach based on mfcc and hmm, in: Intelligent Engineering Systems (INES), 2012 IEEE 16th International Conference on, IEEE, 2012, pp. 561-566.

 twenty

xviii) S. Kaloudis, D. Anastopoulos, C. P. Yialouris, N. A. Lorentzos, A. B. Sideridis, Insect identification expert system for forest protection, Expert Systems with Applications 28 (3) (2005) 445-452.

xix) A. Henriquez, JB Alonso, CM Travieso, B. Rodríguez-Herrera, F. Bolanos, P. 25 Alpízar, K. Lopez-de Ipina, P. Henriquez, An automatic acoustic bat identification system based on the audible spectrum, Expert Systems with Applications 41 (11) (2014) 5451-5465.

xx) G. Grigg, A. Taylor, H. Me Callum, G. Watson, Monitoring frog communities: an 30 application of machine learning, in: Proceedings of Eighth Innovative Applications of Artificial Intelligence Conference, Portland Oregon, 1996, pp. 1564-1569.

xxi) C.-H. Lee, C.-H. Chou, C.-C. Han, R.-Z. Huang, Automatic recognition of animal vocalizations using averaged mfcc and linear discriminant analysis, Pattern 35 Recognition Letters 27 (2) (2006) 93-101.

xxii) T. S. Brandes, Feature vector se1ection and use with hidden markov models to identify frequency-modulated bioacoustic signals amidst noise, Audio, Speech, and Language Processing, IEEE Transactions on 16 (6) (2008) 1173-1180. 40

xxiii) C.-J. Huang, Y.-J. Yang, D.-X. Yang, Y.-J. Chen, Frog classification using machine learning techniques, Expert Systems with Applications 36 (2) (2009) 3737-3743.

xxiv) MA Acevedo, CJ Corrada-Bravo, H. Corrada-Bravo, LJ Villanueva-Rivera, TM 45 Aide, Automated classification of bird and amphibian calls using machine learning: A comparison of methods, Ecological Informatics 4 (4) (2009) 206-214.

xxv) N. C. Han, S. V. Muniandy, J. Dayou, Acoustic classification of australian anurans based on hybrid spectral-entropy approach, Applied Acoustics 72 (9) (2011) 639-645. fifty

xxvi) W.-P. Chen, S.-S. Chen, C.-C. Lin, Y.-Z. Chen, W.-C. Lin, Automatic recognition of frog calls using a multi-stage average spectrum, Computers & Mathematics with Applications 64 (5) (2012) 1270-1281.

xxvii) C. L. T. Yuan, D. A. Ramli, Frog sound identification system for frog species 5 recognition, in: Context-Aware Systems and Applications, Springer, 2013, pp. 41-50.

xxviii) H. Jaafar, DA Ramli, BA Rosdi, S. Shahrudin, Frog identification system based on local means k-nearest neighbors with fuzzy distance weighting, in: The 8th International Conference on Robotic, Vision, Signal Processing & Power Applications, 10 Springer, 2014, pp. 153-159.

xxix) C. Bedoya, C. Isaza, J. M. Daza, J. D. Lopez, Automatic recognition of anuran species based on syllable identification, Ecological Informatics 24 {2014) 200--209.

 fifteen

xxx) J. Xie, M. Towsey, A. Truskinger, P. Eichinski, J. Zhang, P. Roe, Acoustic classification of australian anurans using syllable features, in: Intelligent Sensors, Sensor Networks and Information Processing (ISSNIP), 2015 IEEE Tenth International Conference on, IEEE, 2015, pp. 1-6.

 twenty

Other examples of bio-acoustic vocalization recognition can be found in the automatic identification of marine mammals where studies on whales stand out. The following publications are examples of this:

xxxi) Mouy, Xavier, Mohammed Bahoura, and Yvan Simard. "Automatic recognition of fin 25 and blue whale calls for real-time monitoring in the St. Lawrence." The Journal of the Acoustical Society of America 126.6 (2009): 2918-2928.

xxxii) Dugan, Peter J., et al. "North Atlantic right whale acoustic signal processing: Part l. Comparison of machine learning recognition algorithms." Applications and Technology 30 Conference (LISAT). 2010 Long Island Systems. IEEE, 2010

xxxiii) Baumgartner, Mark F., and Sarah E. Mussoline. "A generalized baleen whale call detection and classification system." The Journal of the Acoustical Society of America 129.5 (2011): 2889-2902. 35

xxxiv) Seekings, Paul, and John Potter. "Classification of marine acoustic signals using Wavelets & Neural Networks." Proc. of 8th Western Pacific Acoustics Conf. (Wespac8). 2003

 40

There are several patents related to the bio-acoustic identification of species which focus generically on the collection and comparison of data and sound parameters based on their vocalizations. But all of them focus mainly on the identification of birds and none of them contemplate the acoustic identification of reptiles, nor do they take into account their bio-acoustic specificities. In addition, they only contemplate the possibility of identifying 45 non-individual species, subfamilies or genus within a given species. An example of this can be found in the following patents:

xxxv) WO 2005024782 A1 (Wildlife Acoustics Inc, Ian Agranat) "Method and apparatus for automatically identifying animal species from their vocalizations". fifty

xxxvi) US 8599647 B2 (Wildlife Acoustics, Inc.) "Method for listening to ultrasonic animal sounds".

xxxvii) US 7963254 B2 (Pariff Llc) "Method and apparatus for the automatic identification of birds by their vocalizations". 5

xxxviii) US 20130282379 A1 (Tom Stephenson, Stephen Travis POPE) "Method and apparatus for analyzing animal vocalizations, extracting identification characteristics, and using databases of these characteristics for identifying the species of vocalizing animals". 10

xxxix) US 20040107104 A1 (Schaphorst Richard A.) "Method and apparatus for automated identification of animal sounds".

xl) US 8457962 82 (Lawrence P. Jones) "Remote audio surveillance for detection and 15 analysis of wildlife sounds".

As for reptiles, scattered studies focus on the spectrum-temporal analysis of the acoustic characteristics of reptiles, but none of them make use of these characteristics for the automated recognition of these species. In addition, these 20 mainly focus on crocodiles and geckos that are the most communicative species among reptiles.

xli) Vergne, A. L., M. B. Pritz, and N. Mathevon. "Acoustic communication in crocodilians: from behavior to brain." Biological Reviews 84.3 (2009): 391-411. 25

xlii) Wang, Xianyan, et al. "Acoustic signals of Chinese alligators (Alligator sinensis): social communication." The Journal of the Acoustical Society of America 121.5 (2007): 2984-2989.

 30

xliii) Ferrara, Camila R., Richard C. Vogt, and Renata S. Sousa-Lima. "Turtle vocalizations as the first evidence of posthatching parental care in chelonians." Journal of Comparative Psychology 127.1 (2013): 24.

xliv) Labra, Antonieta, et al. Acoustic features of the weeping lizard's distress call. Copeia, 35 2013, vol. 2013, no 2, p. 206-212.

Therefore, it can be observed that there is no record of the automated identification of reptiles through their sound production, both of the species to which it belongs and the individualized monitoring of a specific specimen. The present invention aims at the specific recognition of the species, family, subfamily and genus to which a particular reptile belongs based on its bio-acoustic emission characteristics and by hyper-sizing the fusion transformation of its acoustic characteristics in the brush and temporal domains. Thanks to this step, this solution has not been found in the state of the art, unlike the vocalizations made by 45 other animal species that have vocal cords. This proposal would automatically recognize bio-acoustic vocalizations and emissions of any nature in reptiles. The invention, therefore, would have potential applications in the detection, identification and monitoring of the group of reptile animals (Reptilia) or sauropsida. Thus allowing population control, which in turn has applications in the control of pests or invasive species, in the conservation of species, biological studies of animal behavior, changes in

environmental conditions, etc. Even in the detection of possible pathologies or pests that could affect this animal group. The invention therefore opens a wide range of possibilities of applications in the biological or environmental conservation field. Therefore, its analysis and detection is very important in current and future times.

 5

It is possible to conclude after this background, that the studies that have been developed so far and that have had as characteristic parameter the sounds produced by reptiles, have been used basically for the study of the biological behavior of the species, to characterize the fundamental acoustic parameters of their calls, establish their neurology or study their involvement in their social behavior. Also, the background shows specific work for different species of animals, or general systems based on a classic pattern recognition system, without particularities on how to improve the recognition according to the species or application. The proposed method, unlike what is observed in the state of the art, allows its verbal and nonverbal acoustic parameters to be used to enable the recognition of species 15 by means of a module that increases the hyperdimensionality of the transformation of the applied acoustic characteristics to intelligent systems. This has the advantage of not being invasive, because with a remote microphone system the acoustic signal of the specimens can be captured and analyzed. In addition, it allows the monitoring and detection of these species in conditions of limited visibility. twenty

Summary of the invention

The present invention relates to a method for the identification and census of reptile species from the hyperdimensionality of the transformation of their sound production 25 following the following five steps:

i) Pre-processed acoustic signal emphasizing the regions that contain more information.

 30

ii) Automatic segmentation of verbal and non-verbal calls and sounds detected in the acoustic signal, separating the different sound emissions that may belong to different species or specimens in the audio signal.

iii) Parametric fusion of the characteristics extracted in frequency and time of each segmented sound 35 of the calls or vocalizations to obtain a complete representation of different domains of the sound source.

iv) Transformation of the merged characteristics from generating a hyperdimensionality of them, creating a domain of representation of the more discriminative Markov model.

v) Classification and identification of the species or individual through an automated learning algorithm.

 Four. Five

Description of the figures

Figure 1 schematically details the block diagram of the developed system.

Figure 2 shows the spectrogram shape of the sound emissions of reptiles. fifty

a) Crotalus atrox.

b) Gekko gecko.

c) Alligator mississippiensis. 5

d) Chelonoides nigra.

Figure 3 schematically represents the segmentation of the vocalizations.

 10

a) Calculation of the fast Fourier transform (FFT)

b) Location of the point of greatest energy of the spectrogram

c) The procedure is repeated until the end of the spectrogram. fifteen

Figure 4 schematically details the process of extracting spectral characteristics.

a) Calculation of the fast Fourier transform (FFT). twenty

b) Filtered by means of a Mel triangular filter bank.

c) Transformed Discrete Cosine (DCT).

 25

d) The first 14 OCT coefficients are retained.

Detailed description of a preferred embodiment of the invention

Although the invention is described in terms of a specific preferred embodiment, it will be readily apparent to those skilled in this art that various modifications, redispositions and replacements can be made. The scope of the invention is defined by the claims appended thereto. 35

The proposed invention consists of a method that applies several threads until it reaches the unequivocal identification of the species to which the reptile belongs by means of intelligent systems. The first one performs a preprocessing of the signal (i). Next, a segmentation of the acoustic emissions contained in the audio recording is performed by means of an automatic analysis of its spectrogram (ii). On the audio segments, characteristics in the time and frequency domain are extracted to characterize each verbal or nonverbal sound and all the characteristics are merged to have a robust and unique representation of the sound source. (iii). A transformation of the Markov model will be applied to generate greater dimensionality and new representation of characteristics, over the merged representation. (iv). The transformed parameters are sent to a pattern classification algorithm to perform the identification of the species (v).

Next, the threads listed above are described in detail.

 fifty

(i) The preprocessing of the signal consists of the conversion from stereo to mono of the audio coming from the sound recordings by means of the average between the two channels and the low pass signal is filtered with a cut-off frequency of 18 kHz, because Reptile emissions focus primarily on low frequencies. Next, a pre-emphasis filter is applied to match the energy of the spectrum defined by the equation Y (n) = 5 X (n) - 0.95 * X (n-1), where X (n) is the sound signal and Y (n) the filter outlet. The pre-emphasis filter allows to increase the contribution of the high frequencies in the identification of the specimen.

(ii) Once the signal is preprocessed, the segmentation of vocal or non-vocal sounds is automatically performed by conducting a study of the signal spectrogram. For this, a specially modified version of the Härmä algorithm is used to obtain the segments. For this, the spectrogram is traversed by applying a Hamming window of 11.6 ms duration. and 45% overlap. At each step of the window, the point of greatest energy of the spectrogram is located and the signal is taken to the left and right of that point until the energy drops to 20 dB decibels, repeating the process at each step of the window. Next, an intelligent suppressor of incorrect samples is applied to automatically eliminate those segments that do not contain relevant information for identification. To do this, the dynamic time alignment algorithm in English, Dynamic Time Warping (DTW), has been applied using a mean confidence interval plus 1.8 times the standard deviation. This design, which had not previously been used in the detection of animals, prevents sounds from the natural environment from interfering with the classification process by increasing the recognition success rate.

 25

(iii) Once the different acoustic emissions have been obtained, the spectral characterization coefficients MFCC and LFCC are extracted on each of them, in English `` Mel and Lineal Frequencial Cepstral Coeflcients '', to obtain information on the entire spectrum; and temporary parameters such as the temporal length of the sound and its entropy are obtained.

 30

Next, the set of parameters extracted by each sound is merged; creating a single vector that characterizes both the frequency and time of each call. 14 coefficients are taken for each of the characteristics, thus forming a vector of 28 coefficients for each segment, thus modeling the information of both the high and low frequencies of the sounds produced by reptiles. 35

(iv) A transformation is applied through the use of hidden Markov models, to generate a greater dimensionality of the previous parameterized fusion. This new space of representation will have a greater discrimination and will improve the success rate of recognition, on the classic systems that do not use this type of hyper-dimensioning.

The transformation will allow to move from the vector obtained from the parametric fusion to a vector of much greater dimension, adapted to a representation space that will depend on the number of states and the number of symbols by states of the hidden Markov 45 model (MOM). These vectors will be represented to which the SVM classifier will be applied to obtain a recognition result.

Taking into account the nomenclature used in the description of the MOM classifier, P (X│λ) is interpreted as the probability that a vector of characteristics X (which is the result of the parametric fusion) has been created by the Markov model λ, defined

by the number of states and symbols by state. Then the space adapted for the mentioned mapping of fusion vectors is defined as the gradient of the logarithm of said probability:

 5

Where each component of UX is the derivative with respect to a given MOM parameter and consequently specifies the extent to which each parameter contributes to the parametric fusion vector. In this case, only the derivative has been used with respect to the probability matrix of symbol emission,. Which indicates the probability of 10 emitting a vk symbol while in

image 1

the state j. Where N is the number of states and M the number of symbols per state.

The transformation expression of the Markov model is then obtained by following the expression:

Being δ the Dirac delta function and the gamma matrix Υt (i) an indicative of the probability of being in state i in an instant t. The numerator of the previous expression indicates the number of times each symbol was used in each state.

(v) The vectors are sent to a classification system based on an SVM vector support machine, in English "Support Vector Machine", multi-class identification applying the "One VsOne" strategy that has been previously trained with the audios of the reptile species that you want to identify. Upon leaving the classifier, recognition or detection of the species or individuals in which the study, census or monitoring is desired is obtained. The vector support machine has been configured using a Gaussian type core, K (x, x ') = exp (Υ║x - x'║), with a value of γ = 0.52 and a soft margin of parameter C = 20. The transformation of the Markov model allows a better separation of the sample space! at the entrance of the SVM, separating the different classes in a more efficient way and, therefore, increasing its decision limits, facilitating recognition. This design is, therefore, more effective than classic designs by allowing a more optimal differentiation of the different sounds. The experimental results result in success rates above 99% in the identification of the species to which the reptile belongs. 35

Claims (1)

1. Method for the automatic identification and classification of animal specimens belonging to the Reptilia group (reptiles or sauropsides) by means of the transformation of the Markov model of the parametric fusion of the characteristics of its acoustic bio-5 emissions of verbal and nonverbal sounds , using an intelligent pattern recognition system and an intelligent suppression system for incorrect sounds or segments. The method also identifies the individual, species and gender of each reptile. The detected sounds are passed through an intelligent segment suppressor, to discard the sounds coming from the environment and to all those that contain little or no information for identification. The taxonomic recognition of each reptile is produced by applying the Markov model transformation of the fusion of information of spectral and temporal parameters, which characterize the acoustic emissions of these specimens in frequency and time. In this case, reptile sounds are characterized by:  fifteen (i) The transformation of the Markov model of the parametric fusion of its characteristics to a space of greater dimensionality. The transformation has been calculated using the derivative with respect to the probability matrix of symbol emission; 20 indicating the image 1 probability of issuing a vk symbol while in state j. Where N is the number of states and M the number of symbols per state. The transformation expression of the Markov model is then obtained by following the expression:  25 Being δ the Dirac delta function and the gamma matrix Υt (i) an indicative of the probability of being in state i in an instant t. The numerator of the previous expression indicates the number of times each symbol is used in each state. 30 (ii) Parametric fusion is obtained from the MFCC and LFCC coefficients, in English, "Mel and Lineal Frequencial Cepstral Coeficients", to obtain information on both the low and high frequencies of the reptile's acoustic signals; and of the temporal parameters of bio-acoustic emissions with discriminant characteristics such as entropy and sound duration: Parametric Fusion = {MFCC_Parameters ULFCC_Parameters U Entropy U Sound_Duration}  40 With the new space of representation of greater dimensionality of the transformation of the Markov model of the fusion of parameters, the species or individual is classified and identified by means of an automated learning algorithm.